There’s a quiet assumption behind “AI will know everything”: that everything worth knowing has been written down. It hasn’t. The most decision-relevant knowledge in almost any field lives in people’s heads, never recorded — what a negotiation actually hinged on, why a system fails under load, how a market really buys.
Philosophers call this tacit knowledge — the things we know but cannot easily articulate. It is, almost by definition, absent from the corpus a model trains on. You can read every document a company has ever produced and still not know the one thing its best operator knows.
Why it stays tacit
Some knowledge isn’t written because it’s sensitive. Most isn’t written because it’s hard to write — it’s pattern recognition built from years of cases, surfaced only when a specific situation triggers it. Ask the expert to write it all down and they can’t; ask them the right question and it comes out instantly.
A document is the residue of knowledge. The knowledge itself is still in the person who left it.
The implication for AI systems
If the best knowledge isn’t in the documents, then a system built only on documents has a hard ceiling — and pretending otherwise is how you get confident, wrong answers. The honest design is to use the model for everything that was written down, recognize the edge where the documents run out, and route across that edge to a person.
That’s not a limitation to apologize for. It’s the shape of knowledge itself. The map was never the territory, and the documents were never the expertise.